The Recovery of Risk Preferences from Actual Choices

Abstract

SINCE ITS INTRODUCTION, the expected utility hypothesis has been widely used in the construction of economic models. More recently, attention has focused on the conditions under which it is possible in principle to recover individual investors' risk preferences from their demand for assets (Dybvig and Polemarchakis [2]). This paper represents a first attempt to recover preferences operationally from data on the actual demand for assets. Numerous difficulties are encountered in attempting to measure preferences toward risk in a real world setting. Preferences are revealed through the choices of an individual. But in an uncertain world, these choices also depend on his expectations of future events. Hence, an immediate problem arises in separating the influences of each on such decisions. Problems can also arise in measuring other variables, such as wealth, which influence choices. Because of these difficulties, efforts to classify and measure an individual's risk preferences have been confined to direct assessments in hypothetical environments (e.g. Kahneman and Tversky [4] and Keeney and Raiffa [5, pp. 203-212]).2 In these studies the authors assumed that stated preferences are accurate indicators of actual behavior. The question remains, however, whether individuals actually behave in the way their assessments predict. The purpose of this note is to make some progress in answering this question. In it an experiment is described which infers an individual's risk preferences from his actual choices in a real world environment. Specifically, the risk aversion of a dealer in U.S. Government securities is assessed directly and then estimated statistically from his actual demand for bills in the weekly Treasury auctions. The distribution of returns used in the analysis are calculated from the forecasts made by the dealer himself. In addition to introducing new procedures for measuring preferences, this study provides insights into the reliability of direct assessments in predicting the actual behavior.